AI
Legal Technology
AI for Client Materials in Law Firms – Compare the Options
Compare LexBox Workspace, ChatGPT, Claude, Copilot Cowork, and legal AI services based on storage, model execution, and access to client materials.
LexBox

When a law firm chooses AI for client material, it needs to know where data is stored, where the model is run, and which entities can access the content. Opting out of model training, storing data in a specific region, and deleting material after processing are distinct aspects of data management. This comparison is based on public terms and product documentation reviewed on November 11, 2026.
The assessment applies to the firm's actual agreements, features, and configuration. Base your evaluation on what the mandate requires and verify the terms before using client documents. Feel free to start with the document cheat sheet for law firms. For a broader overview of confidentiality and the firm's responsibilities, read the guide on AI and sensitive client material.
Comparison at a glance
Solution | Suitable for | Avoid or verify |
|---|---|---|
LexBox Workspace with Opper | For lawyers working with client material in disputes. Analyze the entire case file, trace information to the source, and generate drafts using the firm's templates. | Use client material once the firm has approved access to matters, how requests reach the model, data processing, logging, and storage for the selected environment. |
ChatGPT Free, Plus, and Pro | Public or entirely synthetic material. User-controlled data settings. | Keep client data and identifiable matters out of personal accounts. A temporary chat may be retained for a limited time. |
ChatGPT Business | A shared workspace for the firm. Regional storage may be available for primary content. | Avoid if client content must remain within the EU. OpenAI states that copies of prompts and responses may be temporarily stored in the US even when another storage region is selected. |
ChatGPT Enterprise | Managed workspace. Eligible customers can select European storage and regional model execution for covered features. | Verify which features and processing are covered. Web search, apps, and MCP must be evaluated separately. |
Legora | Legal document review and work with client material in the agreed environment. | Verify storage region, where the model runs, sub-processors, and who can obtain support access. |
Harvey | Legal analysis and document work. According to Harvey, model providers do not retain any material after processing. | Check which chats, files, and logs Harvey itself retains, and what applies to the firm's environment. |
Vesence | Legal document work. Files and chats in Workspace are stored according to public terms in the selected primary region. | Some processing may occur elsewhere. Verify the ordered product, the agreement, and how requests reach the enabled models. |
Custom solution on AWS Bedrock | A custom application with the option for model execution within the EU and the ability to prevent the model provider from retaining content from covered requests. | The firm is responsible for permissions, storage, logs, indexes, and backups for its own application. |
Claude Enterprise and Cowork | Document analysis and drafting. In Claude Enterprise, the firm can configure retention periods for conversations and projects. | Check retention times, whether Cowork runs locally or in the cloud, and how external integrations handle client material. Verify for which parts the model provider refrains from retaining material. |
Microsoft 365 Copilot Cowork | Multi-step work with documents and tasks in Microsoft 365 based on user permissions. | Microsoft states that prompts and responses are stored in Sweden when Sweden is the organization's default region. Verify where data is processed and how external integrations handle it. |
Personal and managed ChatGPT subscriptions
ChatGPT Free and Plus
Free and Plus are personal subscriptions with user-managed settings for model training. Temporary chats are not used for training and may be retained for up to 30 days for safety purposes. A practical starting point for the firm is to restrict personal subscriptions to public and entirely synthetic material in accordance with its internal policy.
ChatGPT Pro
Pro falls under the same baseline model for data controls as Free and Plus. Assess personal subscriptions collectively when establishing rules for what information employees are permitted to use there.
OpenAI on data controls for personal subscriptions
ChatGPT Business
By default, OpenAI does not use workspace content for model training. Regional storage is being rolled out and can govern where primary customer content is stored. OpenAI notes that copies of prompts and responses are nevertheless stored in the US for a limited time for safety and abuse monitoring. Regional storage in Business also does not specify where the model is executed.
OpenAI on training and data controls · OpenAI on storage in Business
ChatGPT Enterprise
Eligible customers can select European storage and regional model execution for covered content and features. The regional setting for model execution applies to supported GPU processing. Verify what processing may occur outside the region and what terms apply to, for example, web search, apps, and MCP connections.
OpenAI on storage region and model execution
Legal AI services
Legora
Legora states that customer data is not used to train foundation models and that the platform offers various storage options. Request confirmation for your environment regarding storage region, model execution, support access, and sub-processors.
Legora's security documentation · Legora's security policy
Harvey
Harvey describes regional data storage and requires that model providers do not retain material after processing. Also check which files, chats, and logs Harvey itself retains and what retention policy applies to the contracted environment.
Harvey's security documentation
Vesence
Vesence's public Workspace terms state that chats and files are stored in the selected primary region and that some processing may occur in other locations. The terms also state that customer content is not used for model training without written consent. Verify which product, which data processing agreement, and how requests reach the models in the ordered solution.
Vesence Workspace terms · Vesence's sub-processors
Dedicated AWS environment or governed model access via Opper
Custom solution on AWS Bedrock
In a dedicated AWS account, the firm can configure regional storage and select a configuration where models are run within the EU. Bedrock's setting to avoid data retention can block requests to models that require content to be saved. The application's storage, document indexes, logs, access, and backups need to be configured and verified separately.
AWS on data storage for model requests · AWS on geographic model execution
LexBox Workspace with Opper
LexBox Workspace provides dispute teams with access to legal workflows, the firm's templates, and matters to which the user has permissions via LexBox MCP. Opper is LexBox's partner for model access and offers dedicated hosting as an option. For the selected setup, the firm needs to review where the gateway service and model are executed, and what the client, LexBox, Opper, and the model provider retain. Assess what the model provider stores alongside other storage and logging in the workflow.
LexBox Workspace · Opper Enterprise · Opper's security overview
What does the CLOUD Act mean for a law firm?
US law may require a covered provider to preserve or disclose data in its possession, custody, or control, even when located outside the United States. This concerns disclosure pursuant to applicable legal process. For an organization subject to GDPR, a request from a third-country authority must also be evaluated under applicable GDPR provisions, including Article 48 and the rules on third-country transfers.
Map out who can access stored files, log content, and data in transit or processing. Document which legal entity provides each tier, who controls the encryption keys, and which sub-processors may process the material. A dedicated gateway can control where the intermediary service runs. How requests reach the model, what the client does, and additional storage must be evaluated separately.
US Code, 18 USC § 2713 · EDPB's final guidance on Article 48
Questions to ask before the firm decides
Ask the provider to demonstrate the data flow for a real-world task, such as when a lawyer opens a case file, has the model compare two documents, and exports a draft. Inquire where originals, copies, prompts, responses, and logs reside at each step, who has access to them, and how they are deleted. Reconcile the answers with the firm's guidelines, client terms, and the specific mandate.
The Swedish Bar Association's guidance on generative AI · The Swedish Bar Association on external IT services
START YOUR NEXT MOVE


